Causal structure estimation system, production support system, and program

Abstract

This invention provides a causal structure estimation system, a production support system, and a program with improved accuracy in estimating causal structures. [Solution] A causal structure estimation system 2 comprising: a causal inference processing unit 22 that generates directed edges DE by applying an algorithm 12 to two nodes N included in the data 11; a global inference unit 23 that generates directed edges DE between all pairs of nodes N extracted from a plurality of nodes N; a confidence calculation unit 24 that calculates the confidence of the directed edges DE; a cyclic extraction unit 25 that extracts cyclic groups 41 that form a cyclic state using the directed edges DE generated by the global inference unit 23; a modification unit 26 that deletes the lowest edge LE with the lowest confidence among the directed edges DE between two nodes N constituting the cyclic group 41, or reverses the causal direction of the lowest edge LE; and a sparse processing unit 27 that reduces the number of directed edges DE.

Description

Technical Field

[0001] The present invention relates to a causal structure estimation system, a production support system, and a program.

Background Art

[0002] Conventionally, Patent Document 1 (Japanese Patent Application Laid-Open No. 2006-099482) is known as a technique for estimating a causal structure in analysis target data by structural equation modeling means and generating a non-linear regression model based on the estimated result.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Conventionally, as a method for estimating a causal structure for multivariate data, for example, SCORE is known. This SCORE can perform good estimation for non-linear and additive noise data. However, there is a problem that it is difficult to sufficiently improve the estimation accuracy of the causal structure for data that does not have the above properties, such as data where the noise is not additive. Also, there is a problem that it is difficult to sufficiently improve the estimation accuracy of the causal structure for multivariate data with strong causal relationships, such as when the weights are relatively large.

[0005] The present invention has been made in view of such problems, and aims to provide a causal structure estimation system, a production support system, and a program with improved estimation accuracy of the causal structure.

Means for Solving the Problems

[0006] One aspect of the present invention is Including a storage device and a processor, The aforementioned storage device is Data containing multiple nodes representing multiple variables, An algorithm that does not assume the noise is additive and infers directed edges representing causal relationships between two nodes, and stores the following: The aforementioned processor, A causal inference processing unit that generates the directed edge between the two nodes by applying the algorithm to the two nodes included in the data, A global inference unit that extracts all pairs of nodes from the plurality of nodes, and generates directed edges between all pairs of nodes by having the causal inference processing unit perform processing on each of the extracted pairs of nodes, A confidence calculation unit calculates the confidence level of the directed edge generated by the overall inference unit, A cyclic extraction unit extracts from the plurality of nodes a cyclic group that forms a cyclic state with the directed edges generated by the overall inference unit, A modification unit that removes the lowest edge with the lowest confidence level among the directed edges between the two nodes constituting the cyclic group, or reverses the causal direction of the lowest edge, The causal structure estimation system includes a sparse processing unit that reduces the number of directed edges after processing by the aforementioned modification unit.

[0007] Other aspects of the present invention include: The above causal structure estimation system, An input unit for inputting the aforementioned data, An output unit outputs the quality result of the workpiece, the cause of any abnormality in the workpiece, or the machine setting conditions or machining conditions of the machine tool, based on the input data and the causal graph generated by the processing by the sparse processing unit. It is a production support system equipped with these features.

[0008] Further aspects of the present invention include: A causal inference processing procedure that generates a directed edge between two nodes by applying an algorithm to two nodes in data containing multiple nodes representing multiple variables, without assuming that the noise is additive, and by inferring a directed edge representing a causal relationship between the two nodes. A total inference procedure that generates directed edges between all pairs of nodes by extracting all pairs of nodes from the plurality of nodes and performing the causal inference processing procedure for each of the extracted pairs of nodes, A confidence calculation procedure for calculating the confidence level of the directed edge generated by the overall inference procedure, A cyclic extraction procedure for extracting cyclic groups from among the plurality of nodes that form a cyclic state by the directed edges generated by the overall inference procedure, A correction procedure which involves removing the lowest edge with the lowest confidence level among the directed edges between the two nodes constituting the cyclic group, or reversing the causal direction of the lowest edge, The program causes a computer to perform a sparse processing procedure to reduce the number of directed edges after processing according to the above correction procedure. [Effects of the Invention]

[0009] According to one aspect of the present invention, even if a cyclic state occurs due to the generation of an inappropriate directed edge between two nodes included in the data, the correction unit can resolve the cyclic state. This improves the accuracy of estimating the causal structure of the data.

[0010] According to another aspect of the present invention, based on causal structure data, the quality result of a workpiece, the cause of any abnormality in the workpiece, or the machine setting conditions or machining conditions of a machine tool can be obtained. This makes it possible to improve the quality of workpieces or increase the productivity of workpieces.

[0011] According to still another aspect of the present invention, even if a cyclic state occurs due to the generation of an inappropriate directed edge between two nodes included in the data, the cyclic state can be eliminated by a correction procedure. Thereby, the estimation accuracy of the causal structure for the data can be improved.

Brief Description of the Drawings

[0012] [Figure 1] Block diagram showing a production support system according to Embodiment 1. [Figure 2] Block diagram showing a storage device according to Embodiment 1. [Figure 3] Block diagram showing a processor according to Embodiment 1. [Figure 4] Flowchart showing the processing of the causal structure system according to Embodiment 1. [Figure 5] Schematic diagram showing data according to Embodiment 1. [Figure 6] Schematic diagram showing the state after the processing by the cycle extraction unit in Embodiment 1. [Figure 7] Schematic diagram showing the state after the processing by the correction unit in Embodiment 1. [Figure 8] Schematic diagram showing the state after the processing by the correction unit in Embodiment 1. [Figure 9] Schematic diagram showing the state after the processing by the sparse processing unit in Embodiment 1. [Figure 10] Block diagram showing a processor according to Embodiment 2. [Figure 11] Part of a flowchart showing the processing of the causal structure system according to Embodiment 2. [Figure 12] Part of a flowchart showing the processing of the causal structure system according to Embodiment 2. [Figure 13] Schematic diagram showing data according to Embodiment 2. [Figure 14] Schematic diagram showing the state after the processing by the cycle extraction unit in Embodiment 2. [Figure 15]A schematic diagram showing the state after processing by the modification unit has been performed in Embodiment 2. [Figure 16] A schematic diagram showing the state of data related to a modified example of Embodiment 2 after processing by the cyclic processing unit. [Figure 17] A schematic diagram showing a modified version of Embodiment 2 in which the process of reversing the orientation of a directed edge is repeated. [Figure 18] A schematic diagram showing the state after processing by the modification unit has been performed in a modified example of Embodiment 2. [Figure 19] A schematic diagram showing the state after processing by the modification unit has been performed in a modified example of Embodiment 2. [Figure 20] A schematic diagram showing the state after processing by the modification unit has been performed in a modified example of Embodiment 2. [Modes for carrying out the invention]

[0013] (Embodiment 1) Referring to Figures 1 to 4, the production support system 1 according to this embodiment 1 will be described. As shown in Figure 1, the production support system 1 comprises a causal structure estimation system 2, an input unit 3, an output unit 4, a machine tool 5, and a server 6. The causal structure estimation system 2, the input unit 3, the output unit 4, the machine tool 5, and the server 6 are connected to a network 7 and configured to communicate with each other. However, there may be two or more machine tools 5, or they may be omitted.

[0014] Server 6 is equipped with any storage medium, such as a hard disk drive or semiconductor memory. However, Server 6 may be a virtual server on the internet or an intranet.

[0015] The machine tool 5 performs a predetermined machining operation on a workpiece (not shown). The machine tool 5 is not particularly limited and may include, for example, a lathe, machining center, milling machine, gear cutting machine, boring machine, etc.

[0016] The causal structure estimation system 2 comprises a storage device 10, a processor 20, and a display unit 30. In this embodiment 1, the storage device 10, the processor 20, and the display unit 30 constitute a computer.

[0017] The storage device 10 can be any storage device 10 as appropriate, such as a hard disk drive, semiconductor memory, or USB (Universal Serial Bus) memory. As shown in Figure 2, the storage device 10 stores data 11, algorithm 12, causal graph 13, and program 14. However, the storage device 10 may be configured to be located on the server 6.

[0018] Data 11 includes multiple nodes N representing multiple variables. Data 11 includes machine setting conditions or machining conditions of the machine tool 5 that perform a predetermined process on the workpiece, and the quality result of the workpiece after the predetermined process has been performed on the workpiece. However, Data 11 may also be configured to include data other than those described above.

[0019] Algorithm 12 is an algorithm for estimating causal relationships in data 11. Algorithm 12 does not assume that the noise is additive and infers directed edges DE representing causal relationships between two nodes N. A directed edge DE is defined as an edge that connects two nodes N and has a direction.

[0020] Algorithm 12 is not particularly limited and can be appropriately selected from, for example, Regression Error Based Causal Inference (RECI), Information-Geometric Causal Inference (IGCI), etc.

[0021] Each algorithm 12 is applicable to data 11 to which predetermined assumptions are made. In other words, when applying an algorithm 12 with assumptions corresponding to data 11 having predetermined properties, high-accuracy causal inference can be performed. However, when applying an algorithm 12 with assumptions that do not correspond to the properties of data 11, the accuracy of causal inference may decrease.

[0022] RECI makes assumptions such as the data 11 being twice differentiable. In this embodiment 1, RECI is adopted as algorithm 12.

[0023] IGCI makes assumptions about data 11, such as that it is nonlinear and noise-free.

[0024] The causal graph 13 connects two of the multiple nodes N with a directed edge DE, showing the causal relationship between the multiple nodes N. The causal relationship includes the causal direction, which is the direction of causality. The causal relationship may also include weights.

[0025] Program 14 causes the computer to execute the causal inference procedure (S2), the overall inference procedure (S2), the confidence calculation procedure (S3), the cyclic sampling procedure (S4), the correction procedure (S5, S7), and the sparse processing procedure (S6) (see Figure 4).

[0026] The causal inference processing procedure (S2) generates a directed edge DE between two nodes N by applying an algorithm 12 to two nodes N in data 11 containing multiple nodes N representing multiple variables, without assuming that the noise is additive, and by inferring a directed edge DE representing a causal relationship between the two nodes N.

[0027] The overall inference procedure (S2) extracts all pairs of nodes N from multiple nodes N, and generates directed edges DE between all pairs of nodes N by executing the causal inference procedure (S2) for each of the extracted pairs of nodes N.

[0028] The confidence calculation procedure (S3) calculates the confidence level of the directed edge DE generated by the overall inference procedure (S2).

[0029] The cyclic extraction procedure (S4) extracts cyclic groups 41 from among multiple nodes N that form a cyclic state using directed edges DE generated by the overall inference procedure (S2).

[0030] The correction procedure (S5, S7) removes the lowest edge LE with the lowest confidence level among the directed edges DE between the two nodes N that make up the cyclic group 41, or reverses the causal direction of the lowest edge LE.

[0031] The sparse processing procedure (S6) reduces the number of directed edges DE after processing by the correction procedures (S5, S7).

[0032] As shown in Figure 3, the processor 20 includes a preprocessing unit 21, a causal inference processing unit 22, an overall inference unit 23, a confidence calculation unit 24, a cyclic extraction unit 25, a correction unit 26, and a sparse processing unit 27.

[0033] The preprocessing unit 21 performs preprocessing on the data 11. The preprocessing is not particularly limited, and any preprocessing can be appropriately selected, such as standardization to align the order of multiple data 11, or averaging to calculate the average value of multiple data 11.

[0034] The causal inference processing unit 22 generates a directed edge DE between two nodes N by applying algorithm 12 to the two nodes N contained in the data 11.

[0035] The overall inference unit 23 extracts all pairs of nodes N from multiple nodes N and has the causal inference processing unit 22 perform processing on each of the extracted pairs of nodes N. This generates directed edges DE between all pairs of nodes N.

[0036] The confidence calculation unit 24 calculates the confidence level of the directed edge DE generated by the overall inference unit 23. The directed edge DE generated by the overall inference unit 23 may include directed edge DE with relatively low estimation accuracy, or directed edge DE that were incorrectly estimated despite having no causal relationship. If such low-accuracy directed edge DE or incorrectly estimated directed edge DE are included, the estimation accuracy of the causal graph 13 may decrease. Therefore, in this embodiment 1, the confidence calculation unit 24 calculates the confidence level of the directed edge DE generated by the overall inference unit 23.

[0037] The confidence calculation unit 24 may be configured to calculate the confidence level of all directed edges DE generated by the overall inference unit 23. Alternatively, the confidence calculation unit 24 may be configured to calculate the confidence level of directed edges DE included in the cyclic group 41, which will be described later.

[0038] The reliability is not particularly limited, and any indicator can be appropriately selected. For example, the reliability could be calculated by first calculating the mean squared error (1st mean squared error) when assuming a causal flow from one node N to the other node N, and second calculating the mean squared error (2nd mean squared error) when assuming a causal flow from the other node N to the first node N, and then calculating the reliability based on the first and second mean squared errors.

[0039] The confidence level can be, for example, the absolute value of the difference between the first mean squared error and the second mean squared error. Alternatively, the confidence level can be, for example, the ratio of the first mean squared error to the second mean squared error, or the ratio of the second mean squared error to the first mean squared error.

[0040] The cyclic extraction unit 25 extracts cyclic groups 41 from among multiple nodes N that form a cyclic state using directed edges DE generated by the overall inference unit 23. A cyclic state is defined as a state in which a node starts from one node N, travels along directed edges DE to other nodes N, and then returns to the original node N.

[0041] The method for extracting the cyclic group 41 is not particularly limited, and any method can be appropriately selected. In this embodiment 1, strongly coupled component decomposition was employed.

[0042] When S is a subset of multiple nodes N, S is said to be strongly connected if, for any two nodes N, one node N can reach another node N. Furthermore, S is said to be a strongly connected component if, when any other set of nodes N is added to the strongly connected set S, it no longer becomes strongly connected. The cyclic extraction unit 25 according to this embodiment 1 extracts strongly connected components, each having three or more nodes N as elements, as a cyclic group 41.

[0043] The modification unit 26 removes the lowest edge LE, which has the lowest confidence level, from among the directed edges DE between the two nodes N that constitute the cyclic group 41. This resolves the cyclic state. However, if there are two or more directed edges DE that have the lowest confidence level among the directed edges DE that constitute the cyclic group 41, two or more directed edges DE may be removed as the lowest edge LE.

[0044] The sparse processing unit 27 reduces the number of directed edges DE after processing by the correction unit 26. The method for reducing the number of directed edges DE is not particularly limited, and any method such as Causal Additive Model (CAM), Least Absolute Shrinkage and Selection Operator (Lasso), or Group Lasso can be selected.

[0045] Returning to Figure 1, the display unit 30 displays the causal graph 13 generated by the processing performed by the sparse processing unit 27. By looking at the causal graph 13 displayed on the display unit 30, the operator can understand the causal relationships in the processing of the workpiece by the machine tool 5.

[0046] The display unit 30 can be a known display device such as a liquid crystal display. The display unit 30 may be configured to be installed in a computer, in a machine tool 5, or as a portable terminal carried by an operator.

[0047] The input unit 3 is not particularly limited, and any input device such as a keyboard, mouse, touch panel, joystick, or voice input device can be selected. The operator inputs the data 11 to the causal structure estimation system 2 via the input unit 3.

[0048] The output unit 4 is not particularly limited, and any output device such as an LCD display, tablet terminal, or printer can be selected.

[0049] The output unit 4 outputs the causal graph 13 estimated by the causal structure estimation system 2 and notifies the worker.

[0050] Furthermore, the output unit 4 outputs the quality result of the workpiece, the cause of any abnormalities in the workpiece, or the machine setting conditions or machining conditions of the machine tool 5, based on the input data 11 and the causal graph 13 generated by processing by the sparse processing unit 27. However, the information output by the output unit 4 is not limited to the above.

[0051] For example, if data 11 regarding a malfunction of machine tool 5 is input, the system can output the predicted workpiece quality result due to the malfunction of machine tool 5 based on the input data 11 and the causal graph 13.

[0052] Furthermore, for example, if data 11 regarding an abnormality that occurred in a workpiece is input, the cause of the abnormality that occurred in the workpiece, such as vibration or deterioration of the machine tool 5, can be output based on the input data 11 and the causal graph 13.

[0053] Furthermore, for example, if a desired numerical value is entered for the dimensions of a workpiece, the machine setting conditions or machining conditions for the machine tool 5 for manufacturing a workpiece with the dimensions specified by the entered numerical value can be output based on the entered data and the causal graph 13.

[0054] Next, the processing of this embodiment 1 will be described with reference to Figure 4. Figure 4 shows a flowchart of the processing of the causal structure estimation system 2 according to this embodiment 1.

[0055] When the causal structure estimation system 2 is activated, the processor 20 acquires data 11 from the storage device 10 or the input unit 3 (S1).

[0056] Next, in step S2, the global inference unit 23 extracts all pairs of nodes N from the multiple nodes N. Then, for each of the extracted pairs of nodes N, the global inference unit 23 instructs the causal inference processing unit 22 to generate a directed edge DE between the two nodes N by applying algorithm 12 to the two nodes N contained in the data 11. In this way, the global inference unit 23 generates a directed edge DE between all pairs of nodes N.

[0057] Next, the cyclic extraction unit 25 extracts cyclic groups 41 from among the multiple nodes N that form a cyclic state using directed edges DE generated by the overall inference unit 23 (S3).

[0058] Next, the modification unit 26 determines whether or not a cyclic group 41 exists (S5). If a cyclic group 41 exists (S5:Y), the modification unit 26 removes the lowest edge LE with the lowest confidence level among the directed edges DE between the two nodes N that constitute the cyclic group 41 (S7). After that, the process returns to S4 in Figure 4.

[0059] If patrol group 41 does not exist (S5:N), proceed to step S6.

[0060] In step S6, the sparse processing unit 27 reduces the number of directed edges DE. As a result, the causal structure estimation system 2 generates the causal graph 13.

[0061] With this, the processing of the causal structure estimation system 2 is completed.

[0062] Next, we will explain the operation of the causal structure estimation system 2 with reference to Figures 5 to 9.

[0063] Figure 5 shows the data 11 after the processing of S1 to S2 in Figure 4 has been completed. That is, after all combinations of two nodes N are extracted from the multiple (six in this embodiment 1) nodes N1 to N6 contained in the data 11, algorithm 12 is applied to each of the combinations of two nodes N, thereby generating directed edges DE1 to DE15 between all combinations of two nodes N. Figure 5 shows the six nodes N1 to N6 indicated by letters enclosed in circles, and the 15 directed edges DE1 to DE15 indicated by arrows. However, the number of nodes N contained in the data 11 is not limited to six, but may be one to five, or seven or more.

[0064] Figure 6 shows the state of data 11 after processing S3 to S4 in Figure 4 is completed. That is, the cyclic extraction unit 25 extracts the cyclic group 41, indicated by the dashed line, from among the multiple nodes N. As shown in Figure 6, the cyclic state is formed in the following order: node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N5), directed edge DE2, and node A(N1).

[0065] Furthermore, a cyclic state is formed at node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE3, and node A(N1). Also, a cyclic state is formed at node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2). In this embodiment 1, the cyclic extraction unit 25 employs strongly connected component decomposition, so according to the definition above, the strongly connected component 42 becomes a cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N5), directed edge DE2, and node A(N1). Therefore, the cyclic extraction unit 25 according to this embodiment 1 extracts a cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N5), directed edge DE2, and node A(N1).

[0066] Furthermore, nodes E(N5) and F(N6), enclosed by the dashed-dotted line, are each strongly connected components 42 consisting of one node N. Since they consist of one node N, a cyclic state is not formed. For this reason, the cyclic extraction unit 25 does not extract nodes E(N5) and F(N6) as cyclic groups 41.

[0067] Figure 7 shows the state after the processing of S7 in Figure 4 has been completed. Referring to Figures 6 and 7 together, it can be seen that the directed edge DE3, which was shown in Figure 6, has been deleted in Figure 7. The directed edge DE3 is the lowest-confidence edge LE in the cyclic group 41 in Figure 6. With the deletion of the directed edge DE3, the cyclic group consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE3, and node A(N1) is dissolved.

[0068] However, there exists a cyclic group 41 consisting of node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2).

[0069] Figure 8 shows the state after processing S4, S4:Y, and S7 in Figure 4 has been completed. Referring to Figures 7 and 8 together, the directed edge DE11 that was shown in Figure 7 has been removed in Figure 8. The directed edge DE11 is the lowest-confidence edge LE in the cyclic group 41 in Figure 7. As a result, the cyclic group 41, which consisted of node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2), has been dissolved. Consequently, the cyclic group 41 no longer exists in Figure 8.

[0070] Figure 9 shows the state after the processing of S6 in Figure 4 has been completed. Referring to Figures 8 and 9 together, the directed edges DE4, DE5, and DE14 that were shown in Figure 8 have been removed in Figure 9. As a result, a causal graph 13 is completed with a reduced number of directed edges DE.

[0071] Next, the effects of this embodiment 1 will be described. The causal structure estimation system 2 according to this embodiment 1 includes a storage device 10 and a processor 20. The storage device 10 stores data 11 including multiple nodes N representing multiple variables, and an algorithm 12 that does not assume that noise is additive and infers directed edges DE representing causal relationships between two nodes N.

[0072] The processor 20 comprises a causal inference processing unit 22, an overall inference unit 23, a confidence calculation unit 24, a cyclic extraction unit 25, a correction unit 26, and a sparse processing unit 27.

[0073] The causal inference processing unit 22 generates a directed edge DE between two nodes N by applying algorithm 12 to the two nodes N contained in the data 11. The overall inference unit 23 extracts all combinations of two nodes N from multiple nodes N and generates a directed edge DE between all combinations of two nodes N by having the causal inference processing unit 22 perform the processing for each of the extracted combinations of two nodes N.

[0074] The confidence calculation unit 24 calculates the confidence level of the directed edge DE generated by the overall inference unit 23. The cyclic extraction unit 25 extracts cyclic groups 41 from among multiple nodes N that form a cyclic state using the directed edge DE generated by the overall inference unit 23. The modification unit 26 removes the lowest edge LE, which has the lowest confidence level among the directed edge DE between two nodes N that constitute the cyclic group 41, or reverses the causal direction of the lowest edge LE. The sparse processing unit 27 reduces the number of directed edges DE after processing by the modification unit 26.

[0075] According to this embodiment 1, even if a cyclic state occurs due to the generation of an inappropriate directed edge DE between two nodes N included in the data 11, the correction unit 26 can resolve the cyclic state. This improves the accuracy of causal structure estimation for the data 11.

[0076] The causal structure estimation system 2 according to this embodiment 1 further includes a preprocessing unit 21 that performs preprocessing on the data 11.

[0077] According to this embodiment 1, preprocessing can be performed on the data 11. The preprocessing is not particularly limited, and any preprocessing can be appropriately selected, such as standardization of the data 11 or averaging to calculate the average value of the data 11. By performing preprocessing on the data 11, the accuracy of estimating the causal structure of the data 11 can be improved.

[0078] The algorithm 12 according to this embodiment 1 is Regression Error Based Causal Inference (RECI). This improves the accuracy of causal structure estimation for data 11 where noise is not additive.

[0079] The confidence calculation unit 24 in this embodiment 1 calculates a first mean squared error, which is the mean squared error when assuming a causal flow from one node N to the other node N, and a second mean squared error, which is the mean squared error when assuming a causal flow from the other node N to the first node N. After that, it calculates the confidence level based on the first mean squared error and the second mean squared error. This improves the accuracy of estimating the causal structure for the data 11.

[0080] The causal structure estimation system 2 according to this embodiment 1 further includes a display unit 30 that displays a causal graph 13 generated by processing by the sparse processing unit 27.

[0081] The causal graph 13 generated after processing by the sparse processing unit 27 has unnecessary directed edges DE removed, making it easier for the operator to understand the causal structure. According to this embodiment 1, a causal graph 13 that is easy for the operator to understand can be displayed, thereby improving the operator's work efficiency.

[0082] The data 11 in this embodiment 1 includes machine setting conditions or machining conditions of a machine tool 5 that performs a predetermined process on a workpiece, and data 11 relating to the quality result of the workpiece after the predetermined process has been performed on the workpiece.

[0083] According to this embodiment 1, the causal structure estimation system 2 can be applied to a machine tool 5 that processes a workpiece.

[0084] The production support system 1 according to this embodiment 1 comprises the above-described causal structure system, an input unit 3 for inputting data 11, and an output unit 4.

[0085] Based on the input data 11 and the causal graph 13 generated by processing by the sparse processing unit 27, the output unit 4 outputs the quality result of the workpiece, the cause of any abnormalities in the workpiece, or the machine setting conditions or machining conditions of the machine tool 5.

[0086] According to this embodiment 1, based on the causal graph 13, the quality result of the workpiece, the cause of any abnormality in the workpiece, or the machine setting conditions or machining conditions of the machine tool 5 can be obtained. This makes it possible to improve the quality of the workpiece or the productivity of the workpiece.

[0087] The program 14 according to this embodiment 1 causes the computer to execute a causal inference processing procedure (S2), an overall inference procedure (S2), a confidence calculation procedure (S3), a cyclic sampling procedure (S4), a correction procedure (S5, S7), and a sparse processing procedure (S6). The causal inference processing procedure (S2) generates a directed edge DE between two nodes N by applying an algorithm 12 to two nodes N included in data 11 containing multiple nodes N representing multiple variables, without assuming that the noise is additive, and by inferring a directed edge DE representing a causal relationship between the two nodes N.

[0088] The overall inference procedure (S2) extracts all pairs of nodes N from multiple nodes N, and generates directed edges DE between all pairs of nodes N by performing the causal inference procedure (S2) on each of the extracted pairs of nodes N. The confidence calculation procedure (S3) calculates the confidence of the directed edges DE generated by the overall inference procedure.

[0089] The cyclic extraction procedure (S4) extracts cyclic groups 41 from among multiple nodes N that form a cyclic state using directed edges DE generated by the overall inference procedure. The correction procedures (S5, S7) remove the lowest edge LE with the lowest confidence level among the directed edges DE between two nodes N that constitute the cyclic group 41, or reverse the causal direction of the lowest edge LE. The sparse processing procedure (S6) reduces the number of directed edges DE after processing by the correction procedures (S5, S7).

[0090] (Embodiment 2) Next, Embodiment 2 will be described with reference to Figure 10. Note that, unless otherwise specified, reference numerals used in Embodiment 2 and later that are the same as those used in the previously described embodiments represent the same components as those in the previously described embodiments.

[0091] As shown in Figure 10, the processor 20 according to this embodiment 2 further includes a reversal determination unit 28 that determines whether the lowest edge LE among the directed edges DE between two nodes N constituting the cyclic group 41 has already been reversed. The method for determining whether the lowest edge LE has already been reversed is not particularly limited, and the following method may be adopted. For example, when the lowest edge LE is reversed, the processor 20 assigns 1 to a flag and stores this flag in the memory device 10. As a result, the reversal determination unit 28 can determine whether the lowest edge LE has already been reversed by determining whether the flag stored in the memory device 10 is 1.

[0092] The modification unit 26 according to this embodiment 2 is configured to delete the inverted edge RE if the lowest edge LE is an inverted edge RE that has already been inverted, and to invert the uninverted edge UE if the lowest edge LE is an uninverted edge UE that has not been inverted.

[0093] Next, the processing of the causal structure estimation system 2 according to this embodiment 2 will be described with reference to Figures 11 and 12. Figure 11 shows a portion of the flowchart of the causal structure estimation system 2 according to this embodiment 2. Since steps S1 to S6 in Figure 11 are the same as steps S1 to S6 in Figure 4, redundant explanations will be omitted.

[0094] In step S5, if a patrol group 41 exists (S5:Y), S11 in Figure 12 is executed. In step S11, the inversion determination unit 28 determines whether the lowest edge LE has already been inverted.

[0095] If the lowest edge LE has already been inverted (S11:Y), the modification unit 26 deletes the already inverted edge RE (S12). Then, the process returns to S4 in Figure 11.

[0096] If the lowest edge LE is not already inverted (S11:N), the correction unit 26 inverts the uninverted edge UE (S13). Then, the process returns to S4 in Figure 11.

[0097] With the above steps completed, the processing of the causal structure estimation system 2 according to this second embodiment is finished.

[0098] Next, the operation of this second embodiment will be explained with reference to Figures 13 to 15. Figure 13 shows the state of data 11 after processing S1 to S2 in Figure 11 has been performed on the data 11. That is, after all pairs of node N are extracted from the multiple nodes N contained in data 11, algorithm 12 is applied to each of the pairs of node N, thereby generating directed edges DE between all pairs of node N. Compared with Figure 5 of the first embodiment, the directions of the directed edges DE2, DE3, DE6, DE11, and DE14 in this second embodiment are different from those in the first embodiment.

[0099] In Figure 14, a cyclic group 41 consisting of node D(N4), directed edge DE13, node E(N5), directed edge DE15, node F(N6), directed edge DE14, and node D(N4) is shown by a dashed line.

[0100] Figure 15 shows the state after the processing of S13 in Figure 12 has been completed. Referring to Figures 14 and 15 together, it can be seen that the directed edge DE14 shown in Figure 14 has its orientation reversed in Figure 15. The directed edge DE14 is the lowest edge LE in the cyclic group 41 in Figure 14. Also, the directed edge DE14 is the unreversed edge UE. By reversing the directed edge DE14, the cyclic group 41 consisting of node D(N4), directed edge DE13, node E(N5), directed edge DE15, node F(N6), directed edge DE14, and node D(N4) can be resolved. After that, sparse processing (S6) is executed. The sparse processing (S6) is the same as in Embodiment 1, so a redundant explanation is omitted. With this, the operation of Embodiment 2 is completed.

[0101] Next, the effects of this second embodiment will be described. The processor 20 according to this second embodiment further includes a reversal determination unit 28 that determines whether the lowest edge LE among the directed edges DE between two nodes N constituting the cyclic group 41 has already been reversed. The modification unit 26 is configured to delete the reversed edge RE if the lowest edge LE is a reversed edge RE that has already been reversed, and to reverse the unreversed edge UE if the lowest edge LE is an unreversed edge UE that has not been reversed.

[0102] If the correction unit 26 performs a process to invert the lowest edge LE, this inversion of the lowest edge LE may result in the creation of a new cyclic group 41. In this case, the lowest edge LE with the lowest reliability in the new cyclic group 41 may be the inverted edge RE that has already been inverted by the correction unit 26. Then, there is a concern that the correction unit 26 may invert this inverted edge RE again. As a result, the process of inverting a specific lowest edge LE may be repeated.

[0103] Therefore, in this second embodiment, the inversion determination unit 28 determines whether the lowest edge LE among the effective edges constituting the cyclic group 41 has already been inverted. Subsequently, the modification unit 26 deletes the inverted edge RE if the lowest edge LE is an already inverted edge RE. This prevents the process of inverting a specific lowest edge LE from being repeated.

[0104] (Modified version of Embodiment 2) Next, the operation of a modified example of Embodiment 2 will be described with reference to Figures 16 to 20. Note that redundant explanations of the same configuration as Embodiment 2 will be omitted. Also, the data 11 of this modified example is the same as that in Figure 5 of Embodiment 1. As shown in Figure 5, a cyclic state is formed in the following order: node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1).

[0105] Furthermore, a cyclic state is formed at node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE3, and node A(N1). Additionally, a cyclic state is formed at node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2).

[0106] The modification unit 26 performs a process to reverse the direction of the directed edge DE3 connecting node A(N1) and node D(N4), as shown in Figure 16. Directed edge DE3 is the lowest-reliable edge LE in the cyclic group 41 in Figure 16. As a result, the cyclic state consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE3, and node A(N1) is resolved. However, the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1) is not resolved.

[0107] In this case, there is concern that the modification unit 26 will further reverse the directed edge DE3 connecting node A(N1) and node D(N4), as shown in Figure 17. Even in this case, the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1) will not be resolved.

[0108] This raises concerns that the modification unit 26 will repeatedly perform the operation of reversing the orientation of the directed edge DE3 connecting node A(N1) and node D(N4).

[0109] In this modified example, the modification unit 26 is configured to delete the inverted edge RE if the lowest edge LE is already an inverted edge RE, and to invert the uninverted edge UE if the lowest edge LE is an uninverted edge UE. As a result, as shown in Figure 18, the directed edge DE connecting node A(N1) and node D(N4) is deleted because it is already an inverted edge RE. Consequently, the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE3, and node A(N1) is dissolved.

[0110] However, the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1) has not been resolved. Also, the cyclic group 41 consisting of node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2) has not been resolved.

[0111] In the cyclic group 41 shown in Figure 18, if the directed edge DE6 is the lowest edge LE, the modification unit 26 may end up repeatedly reversing the direction of the directed edge DE6 connecting node C(N3) and node B(N2), just as it did with DE3. For this reason, the modification unit 26 deletes the directed edge DE6 connecting node C(N3) and node B(N2), as shown in Figure 19. This eliminates the cyclic group consisting of node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE6, and node B(N2).

[0112] Next, the modification unit 26 determines whether the directed edge DE11 connecting node C(N3) and node D(N4) has already been reversed in the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1). Since the directed edge DE11 connecting node C(N3) and node D(N4) is an unreversed edge UE, the modification unit 26 reverses the directed edge DE11 connecting node C(N3) and node D(N4), as shown in Figure 20. As a result, the cyclic group 41 consisting of node A(N1), directed edge DE1, node B(N2), directed edge DE7, node D(N4), directed edge DE11, node C(N3), directed edge DE2, and node A(N1) is resolved. Subsequently, sparse processing (S6 in Figure 11) is performed. This completes the operation of this modified example.

[0113] The present invention is not limited to the embodiments described above, and can be applied to various embodiments without departing from its spirit. [Explanation of Symbols]

[0114] 1: Production support system, 2: Causal structure estimation system, 3: Input unit, 4: Output unit, 5: Machine tool, 10: Memory device, 11: Data, 12: Algorithm, 13: Causal graph, 14: Program, 20: Processor, 21: Preprocessing unit, 22: Causal inference processing unit, 23: Overall inference unit, 24: Confidence calculation unit, 25: Cyclic extraction unit, 26: Correction unit, 27: Sparse processing unit, 28: Inversion judgment unit, 30: Display unit, 41: Cyclic group, DE, DE1~DE15: Directed edge, LE: Lowest edge, N, N1~N6: Node, RE: Inverted edge, S2: Causal inference processing procedure, Overall inference procedure, S3: Confidence calculation procedure, S4: Cyclic extraction procedure, S5: Correction procedure, S6: Sparse processing procedure, S7: Correction procedure, UE: Uninverted edge

Claims

1. Including a storage device and a processor, The aforementioned storage device is Data containing multiple nodes representing multiple variables, An algorithm that does not assume the noise is additive and infers directed edges representing causal relationships between two nodes, and stores the following: The aforementioned processor, A causal inference processing unit that generates the directed edge between the two nodes by applying the algorithm to the two nodes included in the data, A global inference unit that extracts all pairs of nodes from the plurality of nodes, and generates directed edges between all pairs of nodes by having the causal inference processing unit perform processing on each of the extracted pairs of nodes, A confidence calculation unit calculates the confidence level of the directed edge generated by the overall inference unit, A cyclic extraction unit extracts from the plurality of nodes a cyclic group that forms a cyclic state with the directed edges generated by the overall inference unit, A modification unit that removes the lowest edge with the lowest confidence level among the directed edges between the two nodes constituting the cyclic group, or reverses the causal direction of the lowest edge, A causal structure estimation system comprising: a sparse processing unit that reduces the number of directed edges after processing by the aforementioned modification unit.

2. The aforementioned processor further, The system includes a reversal determination unit that determines whether the lowest edge among the directed edges between the two nodes constituting the circulating group has already been reversed, The aforementioned modification section is, If the minimum edge is already an inverted edge, the inverted edge is deleted. The configuration is such that if the minimum edge is an un-reversed edge, the un-reversed edge is reversed. The causal structure estimation system according to claim 1.

3. Furthermore, the causal structure estimation system according to claim 1 further comprises a preprocessing unit that performs preprocessing on the aforementioned data.

4. The causal structure estimation system according to claim 1, wherein the algorithm is Regression Error Based Causal Inference (RECI).

5. The confidence calculation unit is, For the two nodes mentioned above, the first mean squared error is calculated, which is the mean squared error assuming a causal flow from one node to the other node, and the second mean squared error is calculated, which is the mean squared error assuming a causal flow from the other node to the first node. The confidence level is calculated based on the first mean square error and the second mean square error. The causal structure estimation system according to claim 1.

6. Furthermore, the causal structure estimation system according to claim 1 further comprises a display unit that displays a causal graph generated by the processing by the sparse processing unit.

7. The causal structure estimation system according to claim 1, wherein the data includes machine setting conditions or machining conditions of a machine tool that performs a predetermined process on a workpiece, and data relating to the quality result of the workpiece after the predetermined process has been performed on the workpiece.

8. The causal structure estimation system according to claim 7, An input unit for inputting the aforementioned data, An output unit outputs the quality result of the workpiece, the cause of any abnormality in the workpiece, or the machine setting conditions or machining conditions of the machine tool, based on the input data and the causal graph generated by the processing by the sparse processing unit. A production support system equipped with the following features.

9. A causal inference processing procedure that generates a directed edge between two nodes by applying an algorithm to two nodes in data containing multiple nodes representing multiple variables, without assuming that the noise is additive, and by inferring a directed edge representing a causal relationship between the two nodes, A total inference procedure that generates directed edges between all pairs of nodes by extracting all pairs of nodes from the plurality of nodes and executing the causal inference procedure for each of the extracted pairs of nodes, A confidence calculation procedure for calculating the confidence level of the directed edge generated by the overall inference procedure, A cyclic extraction procedure for extracting cyclic groups from among the plurality of nodes that form a cyclic state by the directed edges generated by the overall inference procedure, A correction procedure which involves removing the lowest edge with the lowest confidence level among the directed edges between the two nodes constituting the cyclic group, or reversing the causal direction of the lowest edge, A program that causes a computer to perform a sparse processing procedure to reduce the number of directed edges after processing according to the above correction procedure.

Images